evaluate using one task for regression analysis

[Archaeopteryx]

At the moment, every 10 minutes a decision task is executed which schedules 15 (arbitrary number) child tasks, one for each recent autoland revision, to do regression analysis for CI tasks. The data requested from Treeherder is cached per task and the requests to group_results of Treeherder are relatively expensive.

If the analysis is done in one task, the data is requested once (or the cache hit) but the execution time might be longer.

@La0 What do you think about this? Do you want to pick it up next?

[marco-c]

There is a shared cache between tasks (

mozci/mozci/console/commands/decision.py

Line 138 in 2ea2583

f"mozci-classifications-{environment}": "/cache",

), the problem is that not all tasks run on the same worker, and the cache is shared by worker.

We could either use a single task for everything, or try a shared cache that isn't only shared by worker (e.g. caching on S3 with

mozci/mozci/util/cache_stores.py

Line 160 in 2ea2583

class S3Store(Store):

).

[marco-c]

Once we fix the perf issues, we'll have to revert taskcluster/community-tc-config#951.